tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
efb1356 verified | # Cohere2: causal language modeling with Transformers | |
| This page is the Transformers contract of the Cohere2 MIC and also a usage example. [doc-builder](https://github.com/huggingface/doc-builder/blob/main/docs/runnable-code-blocks.md) renders the `runnable` blocks below as documentation, and its pytest plugin runs them as written against `CohereLabs/tiny-aya-global`. The MIC runner runs the same blocks against the slim fixture and the checkpoint in [`integration.json`](../../integration.json). Lines marked `# doc-builder: hide` only matter there: they swap in the pinned checkpoint or variant and print the result record. | |
| ## Predict the next token | |
| ```py runnable:causal_lm | |
| import torch | |
| from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed | |
| import json # doc-builder: hide | |
| import os # doc-builder: hide | |
| model_id = "CohereLabs/tiny-aya-global" | |
| model_id = os.environ.get("MIC_MODEL", model_id) # doc-builder: hide | |
| variant = json.loads(os.environ.get("MIC_MODEL_KWARGS", "{}")) # doc-builder: hide | |
| device = "cuda" if torch.cuda.is_available() else "cpu" | |
| set_seed(0) # seeds torch, NumPy, and Python: loading, dropout, and sampling draw random numbers | |
| tokenizer = AutoTokenizer.from_pretrained(model_id) | |
| # The checkpoint stores bfloat16 weights; load in float32 so the training step below is numerically stable. | |
| model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.float32).to(device) | |
| if variant: # doc-builder: hide | |
| del model # doc-builder: hide | |
| model = AutoModelForCausalLM.from_pretrained(model_id, **{"dtype": torch.float32, **variant}).to(device) # doc-builder: hide | |
| # Left padding keeps the last position the next-token probe for every row. | |
| tokenizer.pad_token = tokenizer.eos_token | |
| tokenizer.padding_side = "left" | |
| prompts = [ | |
| "The capital of France is", | |
| "Write a short sentence about the ocean:", | |
| ] | |
| inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(device) | |
| model.eval() | |
| with torch.no_grad(): | |
| logits = model(**inputs).logits | |
| # With left padding the next token is read at the last position, never position 0. | |
| next_logits = logits[:, -1, :] | |
| next_token_ids = next_logits.argmax(dim=-1) | |
| print(tokenizer.batch_decode(next_token_ids.unsqueeze(-1))) | |
| ``` | |
| ## One training step | |
| ```py runnable:causal_lm:2 | |
| # Train on labels the model does not predict: shift the token ids by one modulo the vocabulary. | |
| labels = (inputs["input_ids"] + 1) % model.config.vocab_size | |
| labels[inputs["attention_mask"] == 0] = -100 | |
| model.train() | |
| optimizer = torch.optim.SGD([p for p in model.parameters() if p.requires_grad], lr=0.01) | |
| loss = model(**inputs, labels=labels).loss | |
| loss.backward() | |
| before = {n: p.detach().cpu().clone() for n, p in model.named_parameters() if p.requires_grad} # doc-builder: hide | |
| optimizer.step() | |
| print(f"loss: {loss.item():.4f}") | |
| if os.environ.get("MIC_RUNNER"): # doc-builder: hide | |
| named = dict(model.named_parameters()) | |
| checked = ( | |
| "model.layers.0.self_attn.q_proj.weight", | |
| "model.layers.0.mlp.gate_proj.weight", | |
| "model.norm.weight", | |
| ) | |
| result = { | |
| "runtime": { | |
| "device": model.device.type, | |
| "dtype": str(model.dtype).removeprefix("torch."), | |
| "attn_implementation": model.config._attn_implementation, | |
| }, | |
| "next_token_ids": next_token_ids.tolist(), | |
| "next_token_logits": next_logits[:, next_token_ids].float().tolist(), | |
| "training": { | |
| "loss": loss.item(), | |
| "gradients": { | |
| n: bool( | |
| named[n].grad is not None | |
| and torch.count_nonzero(named[n].grad) > 0 | |
| and torch.isfinite(named[n].grad).all() | |
| ) | |
| for n in checked | |
| }, | |
| "updated": any( | |
| not torch.equal(p.detach().cpu(), before[n]) | |
| for n, p in model.named_parameters() | |
| if n in before | |
| ), | |
| }, | |
| } | |
| print(json.dumps(result)) | |
| ``` | |